PropertyValue
is nif:broaderContext of
nif:broaderContext
is schema:hasPart of
schema:isPartOf
nif:isString
  • We used hospital discharge data in a retrospective multilevel analysis of individual risk of cesarean section across hospitals. Data for this analysis were de-identified, and as such, the study was granted exemption from review by the University of Minnesota Institutional Review Board (study number 1011E92980). We used data from the 2009 and 2010 Nationwide Inpatient Sample (NIS) from the Healthcare Cost and Utilization Project (HCUP) by the Agency for Healthcare Research and Quality (AHRQ). The NIS is an all-payer inpatient claims database designed to approximate a 20% stratified sample of US hospitals [25]. While it contains only administrative data, not clinical information, it is one of the most comprehensive national sources of information on hospital-based care in the US and has been regularly used in health services research [26]–[28]. The NIS is designed to approximate a 20% sample of all US community hospitals (non-federal, short-term, general, and other specialty hospitals, including obstetrics-gynecology, ear-nose-throat, orthopedic, and pediatric institutions). The sample includes both public hospitals and academic medical centers, but excludes short-term rehabilitation hospitals, long-term non-acute care hospitals, psychiatric hospitals, and alcoholism/chemical dependency treatment facilities. The hospitals in the NIS are identified using five strata (ownership/control, bed size categories defined by AHRQ, teaching status, urban/rural location, and US region) with sampling probabilities proportional to the number of US community hospitals in each stratum. HCUP provides weights that account for survey features to ensure national representativeness. Detailed information on the NIS dataset, methodology, and variables is publicly available (http://www.hcup-us.ahrq.gov/databases.jsp). Our analyses focused on hospitals that reported discharges with neonatal and/or maternal diagnoses and procedures. From these hospitals, we used a validated methodology to identify hospital discharge records for obstetric deliveries [29]. Our final dataset included 1,475,457 births in 1,373 hospitals in 46 states, including 1,241,255 births to mothers with no prior cesarean sections. We calculated prevalence of cesarean delivery among two groups of women: (1) all women and (2) all women with no prior cesarean deliveries. We identified cesarean delivery using International Classification of Diseases, 9th revision (ICD-9) procedure codes (740X, 741X, 742X, 744X, 7499) as well as Diagnosis Related Group payment codes (370, 371), consistent with validated methods and prior research using the HCUP NIS data [29],[30]. We identified prior cesarean section by ICD-9 codes (65420, 65421, 65423). We calculated the individual likelihood of cesarean at each hospital as the percentage of cesarean sections among deliveries by all women and as the percentage of cesarean sections among deliveries by women with no prior cesarean sections (primary cesarean section) during 2009 and 2010. We also identified, as closely as these data allow, two other groups of women based on their risk status, consistent with AHRQ Inpatient Quality Indicator #33 [31] and used or identified in prior research [17],[19],[32]. These groups are (1) lower risk women, excluding those with preterm delivery (prior to 37 wk gestation; ICD-9 codes 6442, 64420, 64421), multiple gestation (ICD-9 codes 651, 6510X, 6511X, 6512X, 6513X, 6514X, 6515X, 6516X, 6518X, 6519X), fetal malpresentation (ICD-9 codes 652X, 6600X), and prior cesarean delivery, and (2) higher risk women, including those with preterm delivery, multiple gestation, fetal malpresentation, or prior cesarean section. We used a unique hospital identification code to group deliveries by hospital. We also used hospital-specific data on bed size (as defined by AHRQ, accounting for geographic location), teaching status, and rural versus urban location. Hospital teaching status was based on information from the American Hospital Association's Annual Survey of Hospitals. Classification of hospitals as either urban or rural was based on Core Based Statistical Area codes from 2000 census data. Measurement of hospital characteristics replicated previously published studies using HCUP data [19],[27]–[30], and detailed information on each of these data elements is available on the HCUP website (http://www.hcup-us.ahrq.gov/databases.jsp). We also included fixed effects for state in fully adjusted models. Individual-level covariates are based on administrative records, ICD-9 diagnosis and procedure codes, and Clinical Classifications Software (CCS) codes, developed by HCUP for use with ICD-9 codes. Covariates include maternal age, race/ethnicity, and insurance status (primary payer: private insurance, Medicare, Medicaid, self-pay/uninsured, or other), and maternal and infant medical conditions, including diagnoses of the following complications of pregnancy, labor, and delivery: diabetes in pregnancy (both diabetes mellitus and gestational diabetes; ICD-9 codes 6488XX, 250XX), hypertension in pregnancy (including pre-eclampsia and eclampsia; ICD-9 codes 6420X, 6421X, 6422X, 6423X, 6424X, 6425X, 6426X, 6424, 6425, 6426, 6426XX), hemorrhage during pregnancy or placental complications (including placenta previa and placenta accreta; CCS code 182), fetal disproportion or obstruction of labor (CCS code 188), and fetal distress (CCS code 190). Race/ethnicity is self-reported; specific response categories vary by state but are harmonized by HCUP into the following mutually exclusive categories: black, white, Hispanic, Asian, Native American, and other [25],[26]. In this study, race/ethnicity is included as a factor connected to cultural preferences and practices regarding childbirth. Hospital cesarean section rates for each of the four groups described above (all women, all women with no prior cesarean, lower risk women, and higher risk women) were graphed using funnel plots, by the number of annual deliveries in each risk group. Funnel plots show outcomes in the context of precision, demonstrating how the institution performs compared to control limits (in this case, the 99% prediction interval around the calculated mean) [33]. The data structure for the analysis was hierarchical, with births (n = 1,475,457) at level 1, and hospitals (n = 1,373) at level 2 [34]. We used multilevel logistic regression models to quantify how much of the variation in cesarean section risk was attributable to hospitals. First we fit null models to describe the overall variation in cesarean section risk across hospitals for all births, and for births to women without prior cesarean sections. If the distribution of individuals with more medical complications caused hospital differences in the likelihood of cesarean section, we would expect to see less hospital-level variability in cesarean section use among the lower risk population than among the overall population, as measured by nonoverlapping credible intervals around the hospital variance estimate for these two groups. We then extended the null models to include maternal age, race/ethnicity, insurance status, and individual clinical diagnoses at level 1, and hospital bed size and location/teaching status and state at level 2. If individual clinical diagnoses—that is, medical conditions meeting diagnostic criteria—are driving hospital differences in the likelihood that an individual woman has a cesarean section, we would expect that accounting for clinical diagnoses and for hospital variables associated with greater resources or higher risk patients would reduce any hospital-level variability observed under null models. A significant reduction in the hospital variance (as indicated by nonoverlapping credible intervals after covariate adjustment) would suggest that hospital differences largely reflect the clustering of demographic and/or medical conditions of individuals by hospital. We tested a range of specifications for individual clinical diagnoses, and results were robust to these sensitivity analyses. Missing data were less than 5% for all variables except race/ethnicity (13%) and were handled using conventional methods in multilevel models. Data management tasks were conducted using SAS version 9.2. We used Markov chain Monte Carlo methods to fit Bayesian analytic models, where distributions for the model parameters were first estimated with predictive quasi-likelihood approximation with a second-order Taylor linearization procedure as implemented in MLwiN version 2.1 [35]. Bayesian models used a Metropolis-Hastings sampling algorithm, with the first 500 iterations dropped as burn in, and a chain of 5,000 iterations.
rdf:type